The Search for Genetic Modifiers of Disease Severity in the β-Hemoglobinopathies

Abstract

Sickle cell disease (SCD) and β-thalassemia, two monogenic diseases caused by mutations in the β-globin gene, affect millions of individuals worldwide. These hemoglobin disorders are characterized by extreme clinical heterogeneity, complicating patient management and treatment. A better understanding of this patient-to-patient clinical variability would dramatically improve care and might also guide the development of novel therapies. Studies of the natural history of these β-hemoglobinopathies have identified fetal hemoglobin levels and concomitant α-thalassemia as important modifiers of disease severity. Several small-scale studies have attempted to identify additional genetic modifiers of SCD and β-thalassemia, without much success. Fortunately, improved knowledge of the human genome and the development of new genomic tools, such as genome-wide genotyping arrays and next-generation DNA sequencers, offer new opportunities to use genetics to better understand the causes of the many complications observed in β-hemoglobinopathy patients. Here I discuss the most important factors to consider when planning an experiment to find associations between β-hemoglobinopathy-related complications and DNA sequence variants, with a focus on how to successfully perform a genome-wide association study. I also review the literature and explain why most published findings in the field of SCD modifier genetics are likely to be false-positive reports, with the goal to draw lessons allowing investigators to design better genetic experiments.

Sickle cell disease and β-thalassemia show extreme clinical heterogeneity. Successfully identifying genetic modifiers of disease severity requires careful experimental design.

T wo of the promises of the Human Genome Project were to give the medical community the resources to better grasp interindividual variation in disease and treatment and to uncover previously unknown biology (Lander et al. 2001; Venter et al. 2001). In a recent review article to celebrate the 10-year anniversary of our genome’s sequence, the genomicist Eric Lander highlighted many of the outstanding accomplishments made toward achieving these promises: from evolutionary conservation and non-coding RNAs to landscape of epigenetic marks and maps of genetic variation (Lander 2011). As anticipated, the field of human genetics has prospered on the foundation set by the Human Genome Project, in particular through the discovery and characterization of genetic variation by flagship projects such as the HapMap Project (Altshuler et al. 2005, 2010; Frazer et al. 2007) and the 1000 Genomes Project (1000 Genomes Project Consortium 2010). This has led to novel insights into the selective and recombination forces that have shaped our genome, but most and foremost this has had a dramatic impact on our ability to identify genetic risk factors for complex human diseases. Progress in our theoretical understanding of the genetic variation present in our genome—both in terms of its frequency spectrum and correlated nature—was paved by advances in the development of high-throughput DNA genotyping and sequencing technologies. Together, these new genetic knowledge and technologic capabilities led to the discovery of more than 1000 common DNA sequence variants (so-called single-nucleotide polymorphisms, SNPs) associated with common human diseases and other complex traits by genome-wide association studies (GWAS) (Manolio et al. 2009).

Although most of the attention following the release of the human genome sequence has focused on the genetics of common diseases—which is understandable after decades of unfruitful research—our DNA sequence also helped us make significant progress in identifying genes responsible for rare genetic diseases. More recently, we have seen many studies on various human syndromes that used whole-exome DNA resequencing to find causal genes, often after many years of unsuccessful positional cloning and linkage study efforts (Teer and Mullikin 2010). These successes, together with the convergence of tools initially developed to study complex human diseases, have made us better appreciate that many simple Mendelian genetic diseases are, in fact, all but simple. This is particularly true for the β-hemoglobinopathies sickle cell disease (SCD) and β-thalassemia, which are the most common monogenic diseases in the world (Weatherall and Clegg 2001; Weatherall 2010) and are characterized by a very heterogeneous clinical course between patients with the same mutations, ranging from nearly asymptomatic to life-threatening conditions (Steinberg et al. 2009). The interpatient clinical variability in the β-hemoglobinopathies has sparked a vivid interest in identifying genetic modifiers of severity for these diseases. The rationale for this quest is that such genetic modifiers, if they exist, might lead to more precise (or personalized) prognostic tests, but also guide the development of more specific and effective therapies.

Fetal hemoglobin (HbF) levels and concomitant α-thalassemia are the two best-characterized modifiers of severity in SCD and β-thalassemia. Because these two topics are reviewed in other work (Nienhuis and Nathan 2012; Schechter and Elion 2012), they are not extensively discussed here. Rather, I focus on the main practical questions that one needs to address before undertaking an experiment to find new genetic modifiers for SCD or β-thalassemia. What are the main considerations in terms of phenotype definition, genotype acquisition, data analysis, and result validation for such genetic experiments? Finally, with a focus on SCD, I review the existing literature on genetic modifiers and assess the robustness of previously identified genetic associations, highlight the lessons learned from these earlier studies, and discuss how we might build on these results to design future experiments.

PHENOTYPE CONSIDERATIONS

Phenotype Definitions

Until a few years ago, the main preoccupation of human geneticists was to collect sufficient genetic information in their patient samples to begin the search for genetic risk factors. Nowadays, the emergence of genome-wide genotyping platforms and next-generation DNA sequencers has streamlined what used to be an incredible task: One can now sequence a complete human genome at high coverage in 2 wk for less than $5000. This problem solved, scientists are now redirecting their attention to the importance of the quality and harmonization of the phenotypes that are being studied. Although many phenotypes are straightforward to measure (e.g., HbF levels, white blood cell counts), others are more difficult to ascertain (e.g., pulmonary hypertension), and phenotype definitions are not always consistent, even between clinicians at the same clinic. As one example in the field of complex trait genetics, phenotype heterogeneity has been proposed as one possible reason why it has proven so difficult to understand the genetics of psychiatric diseases, despite a very large genetic contribution (Rasetti and Weinberger 2011). Generally, simpler phenotypes that are reproducible and easy to collect (e.g., height) or endo-phenotypes, which are often quantitative traits related to disease etiology (e.g., cholesterol levels for atherosclerosis), have been studied more successfully than complex clinical end points using genetic strategies. There are also important practical considerations, especially in the context of the β-hemoglobinopathies, that influence the large-scale clinical availability of certain phenotypes when attempting to find genetic associations. Transcranial Doppler (TCD) ultrasonography and magnetic resonance imaging (MRI) results provide robust prognostic tools for cerebral vascular accidents (strokes) in children with SCD (Adams et al. 1992; DeBaun et al. 1995). They would thus appear as ideal phenotypes to identify modifiers of stroke risk in SCD. But these exquisite imaging measurements are often impossible to collect in populations in which the prevalence of SCD is high (e.g., Africa), limiting these studies to a small number of patients, which often leads to false-positive reports (Lohmueller et al. 2003) (see below).

The choice of phenotypes and ascertainment scheme, for instance, selection of a case-control sample as defined by the presence or absence of a particular clinical SCD complication, will also determine how analyses of associations of phenotypes with genotype information must be performed. Both dichotomous and continuous traits can be analyzed using simple statistics (e.g., χ² or Fisher’s exact test, linear or logistic regression), but these analyses need to control for relevant confounding factors or matching variables such as age, sex, or medications. They also need to take into account family structure if related individuals are recruited. For continuous traits, the investigator could focus on patients with the most extreme phenotypes. For example, analyzing the bottom and top fifth percentile of the phenotypic distribution would reduce genotyping or sequencing costs by 90%; this strategy was successfully used by one of the two groups that found the association between the BCL11A locus and HbF levels (Menzel et al. 2007). This is a promising strategy as long as there is evidence that the genetic architecture of the trait is the same in the extremes as in the rest of the distribution. In other words, this approach works if the same genes/alleles control phenotypic variation across the phenotypic distribution, which is usually unknown, although recent work on adult height suggests that it is not always the case (Chan et al. 2011). Finally, although retrospective case-control samples are often easier to assemble, one cannot sufficiently emphasize the virtues of prospective studies. They are more expensive and time-consuming but eliminate many caveats and offer unique advantages to identify and test predictive risk factors. Longitudinal data are also amenable to powerful statistical methodologies, which can become important when the phenotypes are rare. This is certainly one of the main reasons why the large prospective Cooperative Study of Sickle Cell Disease (CSSCD), which was initiated in 1977 in the United States to determine the natural history of SCD from birth to death, has had such a tremendous impact on how we understand and treat SCD (Farber et al. 1985; Gaston et al. 1987).

Heritability

Most importantly, for genetic association studies to be successful, the phenotypes need to have a genetic component. Heritability is the key concept here, but it is often misunderstood (Visscher et al. 2008). It is defined, in the case of narrow-sense heritability (h²), as the ratio of the phenotypic variation explained by additive genetic factors over the total phenotypic variation observed in the population. A low heritability estimate indicates that a large fraction of the phenotypic variation (or disease risk on a liability scale) is due to environmental factors (plus other non-additive genetic effects) and that, consequently, the phenotype is harder to track genetically. Heritability is usually estimated by comparing phenotype concordance among relatives, although new genomic-based methods have been developed that use only distantly related individuals (Visscher et al. 2006; Yang et al. 2011).

There is surprisingly little information available on the heritability of the phenotypes and complications relevant to the severity of the β-hemoglobinopathies. The estimate for the heritability of HbF levels is high (h² ∼ 0.6–0.9), but this was obtained in healthy individuals of European descent (Garner et al. 2000; Pilia et al. 2006). In Jamaica, a study with nine identical pairs of twins with sickle cell anemia showed good concordance for many hematological traits, whereas the clinical complications were often discordant (Weatherall et al. 2005). With a focus on stroke, two groups also found evidence of familial aggregation in SCD patients (Driscoll et al. 2003; Kwiatkowski et al. 2003). Finally, in our analysis of genome-wide genotypes in the CSSCD, we identified 104 pairs of full sibs in which we analyzed phenotypic correlation (Table 1). Correlation coefficients were high for HbF levels, pain crisis rate, and stroke; moderate for acute chest syndrome, leg ulcer, osteonecrosis, and priapism; and low for survival. As the size and clinical quality of the β-hemoglobinopathy cohorts increase, estimating heritability of the complications before trying to find their genetic modifiers is certainly an avenue of research that should be pursued.

Table 1.

Correlation of eight phenotypes in 104 pairs of full sibs from the Cooperative Study of Sickle Cell Disease (CSSCD)

Phenotype	Number of pairs with phenotype available (affected siblings/unaffected siblings)	Correlation coefficient
Acute chest syndrome rate	104	0.166
Fetal hemoglobin levels	84	0.579
Leg ulcer	84 (26/142)	0.183
Osteonecrosis	104 (37/171)	0.265
Pain crisis rate	104	0.683
Priapism (male only)	25 (9/41)	0.218
Stroke	104 (16/192)	0.361
Survival	104 (7/201)	0.024

These estimates must be considered carefully because the sample size is small and they do not take into account important confounding factors, such as socioeconomic status.

GENOTYPE CONSIDERATIONS

GWAS

Before the development of genome-wide genotyping arrays, genetic searches for complex diseases and traits used a combination of linkage- and candidate-gene-based approaches. For reasons that fall outside the scope of this article, linkage is inappropriate to find common genetic variants of small effect on phenotypic variation (Risch and Merikangas 1996), and candidate-gene studies are limited to the known biology. Except for the identification of the association between HbF levels and genetic variation at the HBS1L–MYB intergenic region, first by linkage (Craig et al. 1996; Garner et al. 1998) and subsequently by association study (Thein et al. 2007), these strategies have not proven very useful to find genetic modifiers of severity in the β-hemoglobinopathies.

Genetic association analysis consists of testing whether a specific allele, for example, the A allele at an A/G biallelic SNP, is found more often in cases than controls or whether it is correlated with a continuous trait (Hirschhorn and Daly 2005). Because alleles at different markers may be correlated in the genome, a phenomenon called linkage disequilibrium (LD) (Slatkin 2008), it is possible to test association with a large number of variants by genotyping only a subset of them. Because of LD, GWAS test associations systematically across the whole genome, but for markers that are common in the population of interest (usually SNPs with a minor allele frequency [MAF] >5%; for rarer markers, DNA sequencing is required because of weak LD) (see below).

The first successful GWAS was performed in 2005 and identified the complement factor H (CFH) gene as a risk factor for age-related macular degeneration among a cohort of 96 cases and 50 controls, and genotypes at ∼100,000 SNPs (Klein et al. 2005). Since then, GWAS have increased in size both in terms of the number of participants genotyped and markers tested. For example, in a recent study on body mass index, about 2.5 million SNPs were analyzed in about 250,000 individuals (Speliotes et al. 2010). For the β-hemoglobinopathies, GWAS have so far been limited to the discovery of BCL11A and its role on HbF levels (Menzel et al. 2007; Uda et al. 2008), SCD pain crises (Lettre et al. 2008), and β-thalassemia transfusion dependency (Uda et al. 2008; Nuinoon et al. 2009). In the next sections, the key concepts required for a successful GWAS are reviewed.

Statistical Power

For genetic association analysis, the statistical power is defined as the probability of rejecting the null hypothesis of no association when the null hypothesis is indeed false. In other words, if there is a true association between genotype and phenotype, what is the probability of finding it given your study design? Power depends on sample size, marker allele frequency, linkage disequilibrium (LD) strength between the genotyped markers and the causal variants, effect size of the genetic variants on the phenotype, and the number of independent markers tested. This latter factor, also referred to as the “multiple hypothesis burden,” is particularly important in GWAS because hundreds of thousands or even millions of SNPs are usually tested. To limit the number of false-positive associations (type I errors), we use very stringent significance criteria to declare significance, but this has the disadvantage of reducing power. For GWAS, a generally accepted criterion to claim genome-wide significance is a p-value < 5 × 10⁻⁸ (Altshuler et al. 2005). Power is a critical concept to estimate before embarking on any GWAS project but also essential to interpret published GWAS results correctly. Fortunately, user-friendly tools exist to compute the power of association studies (Purcell et al. 2003; Gauderman and Morrison 2006).

Population: Coverage, Stratification, and Admixture

Genetic variation differs between populations because of mutations, genetic drift, migration, and natural selection. As a consequence, not all genetic variants are present in all populations, or the allele frequency of a given variant may differ across populations. For example, populations of African ancestry are genetically more ancient and carry more genetic variation (both rare and common) than populations of European or Asian descent (1000 Genomes Project Consortium 2010). Because the content of commercial genome-wide genotyping arrays is fixed, they cover more exhaustively the common genetic variations in non-African populations (Bhangale et al. 2005). This can have a major impact on genetic discovery efforts, especially for diseases that are more prevalent in Africans or individuals of African ancestry such as SCD. Novel genotyping arrays that were designed using empirical data from the 1000 Genomes Project, as well as imputation methods (Li et al. 2009) that can probabilistically infer genotypes at ungenotyped markers using genotypes from reference sets such as the HapMap or 1000 Genomes Projects, can, however, mitigate the bias introduced by genotyping the more limited markers present in commercial genotyping arrays.

Allele frequency differences between populations can also result in false-positive genetic association results when the prevalence of a disease (or the distribution of a quantitative trait) differs across these populations. The classic example of this important confounder, called population stratification (Price et al. 2010), is the spurious association between height and the DNA polymorphisms responsible for lactase persistence in populations of European descent because both height and lactose tolerance follow a North–West/South–East cline within Europe (Campbell et al. 2005). Using genome-wide genotypes or a limited set of ancestry informative markers, several methods now exist to detect and correct for the effect of population stratification on association results (Price et al. 2006; Kang et al. 2010). It is also worth mentioning that family-based association tests can be protected from the effect of population stratification, providing additional value to this study design.

Whereas population stratification applies to subpopulations from the same continent (e.g., European Americans), admixture describes individuals whose ancestors are from different continents, such as African Americans or Hispanic Americans (Seldin et al. 2011). The chromosomes in admixed populations are mosaics with chromosomal segments from each ancestral populations: This can be useful for gene mapping but also presents several analytical challenges that must be appropriately accounted for when performing GWAS (Price et al. 2009). Correction for admixture is particularly important in the search for genetic modifiers of the β-hemoglobinopathies because many studied populations are, in fact, admixed.

Replication

As mentioned above, the challenge in GWAS is to identify the few authentic genetic associations among millions of markers tested. We can increase the stringency of the significance threshold to account for the number of tests performed, but this has the disadvantage of decreasing power. To compensate for the power loss, we can analyze more clinical or population samples by GWAS, but this increases costs. Most successful GWAS have used staged designs: All markers are first tested in an initial “discovery” population, and then the most promising markers (based on statistical evidence or biological candidacy) are analyzed in replication cohorts. In fact, the replication of genetic associations in independent samples—ideally, a replication study would test the same phenotype, allele, direction of effect, and genetic model with a different genotyping platform—is now the gold standard and is essentially mandatory before publication because it has been shown to be key to avoid false-positive reports (Lohmueller et al. 2003; Chanock et al. 2007). Given the importance of replicating GWAS findings, any proposal to identify β-hemoglobinopathy modifiers should include a plan to replicate association results in independent samples. It is critical to consider accessibility to replication cohorts (independent samples, same phenotypes) early in the design of any GWAS experiments.

Next-Generation DNA Sequencing

The GWAS framework is extremely efficient to capture the role of common genetic variation (MAF > 5%) in phenotypic variation. Rare (MAF < 0.1%) and low-frequency (MAF 0.1%−5%) variants are usually poorly correlated with common variants surveyed by GWAS and have therefore not been tested comprehensively for their role in disease risk. With the development of affordable next-generation DNA sequencing technologies and data release from the 1000 Genomes Project (1000 Genomes Project Consortium 2010), we now have the tools to test the role of rare and low-frequency variants in complex human phenotypes. Genetic results for HbF, implicating rare familial mutations in the β-globin locus and KLF1 (Borg et al. 2010) and low-frequency variants in MYB (Galarneau et al. 2010), strongly hint that rare genetic variation might also modify disease severity in the β-hemoglobinopathies. With many ongoing whole-exome sequencing projects, we should have a better idea of the role of rare and low-frequency genetic variants in SCD and β-thalassemia complications in the coming years.

REVIEW OF PUBLISHED GENETIC MODIFIERS IN SCD

Several studies have been published describing genetic associations between DNA polymorphisms in strong candidate genes and many complications specific to SCD (for review, see Fertrin and Costa 2010). Table 2 provides the details of most of the published genetic associations with acute chest syndrome, leg ulcer, osteonecrosis, pain crisis, and priapism. Based on these data, we can make the following observations: (1) Most sample sizes are small. (2) Reported odds ratios (OR) are high in comparison to what we now regard as realistic effect sizes in genetic association studies (usually OR < 1.2). (3) For many studies, the significance threshold was not adjusted for the number of tests performed. (4) Few results have been replicated in independent cohorts (Table 2). These are all characteristics usually linked to false-positive reports (Lohmueller et al. 2003), which is reflected by the low discovery power that any of these studies had to find genetic associations even under an ideal scenario (OR = 1.5; last column in Table 2).

Table 2.

Published genetic associations with sickle cell disease–related complications

Complication	References	Sample size (cases/controls)	Gene (variant)	Reported odds ratio	Reported p-value	Appropriate significance thresholda	Replication study (reference)	Power (%)b
Acute chest syndrome	Sharan et al. 2004	18/27	NOS3 (rs2070744)	6.5	0.005	0.05	No	10
Leg ulcer	Nolan et al. 2006	243/516	KL (rs516306)	1.8	0.0076	0.0002 (0.05/215 tests)	No	26
			MAP3K7 (rs157702)	1.6	0.008		No
			SMAD7 (rs736839)	2.0	0.0004		No
	Ofosu et al. 1987	9/29	HLA (B35/Cw4)	17.0	<0.005	0.05	No	10
Osteonecrosis	Baldwin et al. 2005	442/455	KL (rs211239)	2.6	0.001	0.0008 (0.05/66 tests)	Yes, but association did not replicate (Ulug et al. 2009).	47
			BMP6 (rs267196)	1.9	0.001		Yes, but association did not replicate; see text (Ulug et al. 2009).
			ANXA2 (rs7170421)	3.4	<0.001		Yes, but association did not replicate (Ulug et al. 2009).
Pain crisis	Mendonca et al. 2010	48/39	MBL2 (exon 1 0/A and −221 X/Y)	3.2	0.02	0.05	No	29
	Al-Subaie et al. 2009	127/130	GP1BA (rs6065)	1.8	0.004	0.02 (0.05/3 tests)	No	45
			ITGA2B (rs5911)	2.0	0.0008		No
			ITGA2 (rs1801106)	1.9	0.002		No
Priapism	Nolan et al. 2005	148/529	KL (rs2249358)	2.6	Not reported	0.001 (0.05/44 tests)	Yes, but association did not replicate (Elliott et al. 2007).	40
	Elliott et al. 2007	83/116	TGFBR3 (rs7526590)	2.7	0.00058	0.001 (0.05/49 tests)	No	5
			AQP1 (rs10244884)	2.1	0.00068		No
			ITGAV (rs3768780)	2.0	0.0009		No

This is not an exhaustive list. Meeting abstracts were not considered because of the difficulty to assess the quality of the results.

^aThe appropriate significance threshold corresponds to a Bonferroni correction for the number of genes/tests performed in each study.

^bDiscovery power for each study was calculated under the following assumptions: odds ratio = 1.5, effect allele frequency = 20%, sample size = number of cases and controls available in the study, α level = appropriate significance threshold, prevalence as determined from the Cooperative Study of Sickle Cell Disease (CSSCD) phase I clinical data set.

The reported association between SNPs at the BMP6 locus and osteonecrosis risk is a good example of the inconsistency that exists in the literature. The association was first identified in 442 cases and 455 controls from the CSSCD (Baldwin et al. 2005) with a replication attempt in another SCD cohort that included 39 osteonecrosis cases and 205 controls (Ulug et al. 2009). Although the investigators of the replication study claimed that they could replicate the BMP6–osteonecrosis association, careful analysis of the articles shows that the two studies have marginal associations with different BMP6 SNPs (rs267196 and rs3812163) that are 124 kb away from each other and not in LD (r² ∼ 0). By definition, because a different SNP was tested in the replication study, the original association between BMP6 and osteonecrosis has not been replicated yet. Thus, none of the associations presented in Table 2 appear robust.

Stroke

Because of its devastating consequences and apparent heritability (Table 1) (Driscoll et al. 2003; Kwiatkowski et al. 2003), many investigators have attempted to identify genetic risk factors for stroke in SCD (Hoppe et al. 2004, 2007; Sebastiani et al. 2005; Voetsch et al. 2008). In a recent replication study, Flanagan and colleagues attempted to confirm these previous stroke association results in an independent SCD cohort of 130 stroke cases and 103 controls (Flanagan et al. 2011). This study had 44% power to replicate an association with stroke (using the same assumptions reported in Table 2). The researchers could replicate the known protective effect of α-thalassemia on stroke risk, and also reported nominally significant association results for SNPs in four genes: ADYC9, ANXA2, TEK, and TGFBR3 (Flanagan et al. 2011). These results may seem encouraging, but once again they must be interpreted with caution. SNPs in ADYC9, ANXA2, TEK, and TGFBR3 were initially related to stroke risk in SCD through a Bayesian network (Sebastiani et al. 2005). Bayesian networks integrate many variables, in this case, laboratory measures such as HbF levels but also genotypes, to predict disease outcome. Although they are theoretically interesting as methods to analyze large data sets with many variable dependencies, the biological and/or clinical interpretation of the different nodes in the network is difficult. In this specific example, although the Bayesian network was validated in an independent cohort of SCD patients with or without stroke, the marginal effect of the SNPs in the network on stroke risk remains unclear. In fact, we tested by logistic regression in the CSSCD the association between stroke and genotypes at ADCY9-rs2238432, ANXA2-rs11853426, TEK-rs489347, and TGFBR3-rs284875, and observed no significant results (p > 0.15) (KS Lo and G Lettre, unpubl.).

Two Convincing Associations

Besides HbF levels and α-thalassemia, there are two other convincing genetic modifiers of complications in SCD. Importantly, these associations are more compelling because they were first very robustly confirmed with phenotypes in non-SCD populations. Polymorphisms in the promoter of the UGT1A1 gene are associated with serum bilirubin levels in many studies and different populations, including patients with SCD; the same polymorphisms are also associated with the risk of formation of gallstones in SCD patients (Passon et al. 2001; Fertrin et al. 2003; Chaar et al. 2005, 2006; Haverfield et al. 2005; Carpenter et al. 2008; Milton et al. 2012). Renal failure is one of the major complications of SCD and is a strong predictor of mortality in SCD (Platt et al. 1994). Admixture mapping and association studies have identified the MYH9–APOL1 locus as an important genetic risk factor for kidney disease in populations of African ancestry (Kao et al. 2008; Kopp et al. 2008; Genovese et al. 2010). More recently, risk alleles at the MYH9–APOL1 locus were also associated in SCD patients with proteinuria and decreased glomerular filtration rate, both hallmarks of SCD nephropathy (Ashley-Koch et al. 2011). The UGT1A1 and MYH9–APOL1 results are reminders of the importance of large sample size, access to replication cohorts, and clearly defined or well-harmonized phenotypes to find true-positive genetic associations.

CONCLUSIONS

The identification of genetic modifiers of disease severity for the β-hemoglobinopathies promises to improve patient management and treatment. As highlighted in this article, many of the previous associations between SCD complications and genetic variation are false-positive reports. This result is not specific to SCD because most of the association studies performed before the GWAS era were underpowered. Fortunately, robust associations for many complex human phenotypes, including some relevant to hemoglobin disorders (e.g., HbF levels, kidney disease), now exist and allow us to identify the key factors that lead to successful genetic searches: clear phenotype definition and heritability estimate, large sample size and knowledge of statistical power, control for the multiple hypothesis burden and other confounders such as population stratification and admixture, and access to independent samples for replication. Because the importance of these factors are well appreciated and with large collaborations of geneticists and clinicians working on the β-hemoglobinopathies already in place, we are finally in a position to ask convincingly important questions, test clinically meaningful hypotheses, and, hopefully, gain insight into the pathophysiology of SCD and β-thalassemia. This is clearly an exciting time for β-hemoglobinopathy patients, their families, and physicians.